Atomic action recognition and activity analysis focusing on meeting room scenarios

Abstract

While human action recognition is being paid more and more attention during recent years, the accuracy increased greatly. For instance, the two-stream method [1] and LSTM based method [2] all give satisfying accuracy on normal human actions. However, when it turns to atomic actions that can happen simultaneously, the results are not very convincing. Actions that happened in a meeting room scenario are usually atomic so that it would be an ideal environment to study the action recognition model. This report focuses on the procedure on exploring possible methods on atomic action recognition and also the improvements on existing models to make it perform better on meeting room scenarios. The project covered by the report consists of two parts. The first part is to improve a model to fit the meeting room scenario. The second part is to find a method using the model to extract information from or analyzing meeting room videos to check if all the participants are paying attention.